Migration and Consumption Insurance in Bangladesh

Transcription

1 Migration and Consumption Insurance in Bangladesh Costas Meghir (Yale) Mushfiq Mobarak (Yale) Corina Mommaerts (Wisconsin) Melanie Morten (Stanford) October 18, 2017

2 Seasonal migration and consumption insurance Income very volatile in developing country agricultural households Often seasonal component Seasonal and temporary migration is common Informal risk sharing common and important How do these insurance mechanisms interact?

3 This paper Experimental intervention to increase migration Treatment effect of migration on risk sharing Risk sharing improves Is this consistent with limited commitment risk sharing? Estimate model using control data Estimate out-of-sample predictions of the experiment Model matches treatment effects In progress: further counterfactuals

4 Contributions Bryan, Chowdhury and Mobarak (2014) Study effects of migration on migrants We study spillover efffects on risk sharing network Morten (2017) Focus on seasonal migration Exploit experimental variation

5 Key related literature Rural-urban migration Bryan, Chowdhury and Mobarak (2014); Sjaastad (1962); Harris and Todaro (1970) Limited commitment risk sharing Krueger and Perri (2010); Ligon, Thomas, Worrall (2002); Kocherlakota (1996) Interaction between substitute risk management strategies Mobarak and Rosenzweig (2014), Morten (2017) Risk sharing and permanent migration Banerjee and Newman (1998); Munshi and Rosenzweig (2009) Combining structural models and experimental data Kaboski and Townsend (2011), Todd and Wolphin (2006)

6 Outline 1. Data and experiment 2. Reduced form test 3. Model of endogenous risk-sharing and migration

7 Experimental setup Bryan, Chowdhury and Mobarak (2014) North-West Bangladesh (Rangpur division): Population Rangpur: 9.6m; 5.3 m below poverty line Lean season (Monga) prior to Aman rice harvest (Sep-Nov) But, low levels of seasonal out-migration Seasonality of consumption Experiment: August 2008 Small (approx bus ticket) cash or credit incentive Baseline: July 2008 pre-monga Follow up surveys: Nov 08; Nov 09; July 11; Dec villages over two districts Cash: 37; Credit: 31; Information: 16; Control: 16

8 Summary of main experimental results 1. Seasonal/circular migration Increase of 22 p.p. in migration Control: 36%, Treatment: 58% 2. Own household consumption at origin LATE estimate: increase 30% Table 3. Re-migration One year later (no incentives): Mig + 9%; Cons + 28% 2.5 years later: Mig +7%; Cons + 30% 4.5 years later: Mig + 7%; Cons + 35%; Summary stats

9 Why would experiment affect risk sharing? Experiment made migration easier Reduced travel cost Interaction with risk sharing Increases value of outside option (-) Household uses migration as self-insurance, risk-sharing less valuable If migration itself risky, risk-sharing may facilitate (+) Network can help insure risky decision May help insure aggregate shocks (+) Net effect on insurance ambiguous

10 Outline 1. Data and experiment 2. Reduced form test 3. Model of endogenous risk-sharing and migration

11 Testing for risk sharing: Townsend test Standard Townsend test: log c ivt = β log y ivt + γ i + γ vt + ɛ ivt Full risk sharing: β = 0 No risk sharing: β = 1

12 Table: Consumption smoothing among control villages (1) (2) Log total consumption Log food consumption Log income (0.015) (0.014) Observations R-squared

13 Does migration cause risk sharing to get better? Interact treatment with income Intepreting sign of interaction Negative: cons, income less correlated: r/s Positive: cons, income more correlated: r/s log c ivt = β 1 log y ivt + β 2 log y ivt T v + γ i + γ vt + ɛ ivt

14 Table: Effect of migration incentives on consumption smoothing Log total consumption Log food consumption (1) (2) (3) (4) (5) (6) Overall treatment effect ** ** (0.020) (0.019) Group restrictions Unassigned group ** ** (0.024) (0.023) Self-formed group (0.028) (0.029) Assigned group * * (0.030) (0.027) Destination restrictions Unassigned destination ** ** (0.022) (0.023) Assigned destination (0.025) (0.023) Observations R-squared Savings

15 Table: Effect of migration incentives on consumption smoothing, non-migrant sample Log total consumption Log food consumption (1) (2) (3) (4) (5) (6) Overall treatment effect * ** (0.026) (0.023) Group restrictions Unassigned group ** ** (0.030) (0.029) Self-formed group (0.038) (0.037) Assigned group * ** (0.037) (0.032) Destination restrictions Unassigned destination ** * (0.030) (0.029) Assigned destination (0.032) (0.029) Observations R-squared

16 Direct evidence Table: Treatment effect on financial assistance from and to others Would help you Would help you Would ask you for help Would ask you for help and you d ask and you d help Family 0.047* 0.043* 0.111*** 0.106*** (0.026) (0.026) (0.033) (0.031) Control mean [0.730] [0.707] [0.516] [0.475] Friends 0.081*** 0.073** 0.096*** 0.090*** (0.031) (0.030) (0.029) (0.027) Control mean [0.258] [0.239] [0.207] [0.182] Other villagers 0.069** 0.070** 0.106*** 0.105*** (0.028) (0.027) (0.031) (0.026) Control mean [0.628] [0.588] [0.365] [0.306] NGOs 0.067** 0.071** (0.030) (0.029) Control mean [0.540] [0.494] Moneylenders (0.021) (0.020) Control mean [0.208] [0.180] Migrant subsample Non-migrant subsample

17 Risk sharing improved in treatment villages Correlation between income and consumption decreased 4% Food consumption: 4% Effect consistent when look only at non-migrants 5% Food consumption: 5% Suggests risk sharing improved in treatment villages

18 Outline 1. Data and experiment 2. Reduced form test 3. Model of endogenous risk-sharing and migration

19 Move to structural estimation Experiment changes the income process of the village Variance of income Persistence of income Measurement error in income We calibrate LC model using control villages Income process generates consumption stream Estimate model to match risk sharing Validate model out-of-sample with the experiment

20 Limited commitment model Households can walk away from risk sharing model Value of risk sharing needs to be as high as autarky Endogenously incomplete risk sharing Changes in income process affect value of autarky Estimation approach Estimate mig, village income off control villages Then, change migration cost Kocherlakota (1996), Ligon, Thomas and Worrall (2002)

21 AR (1) income process Income process characterized by Variance of measurement error: var(ɛ y ) Variance of persistent shock var(ν) Income persistence ρ log y it = u it + ɛ y it u it = a + ρu it 1 + ν it Estimate separately for treatment and control Identify from cross-person moments Details of moment conditions Naive estimation of model

22 Full model Social planner decides utility now vs promise for future Timing: 1. Village income revealed 2. Migration and contingent utilities chosen 3. Migration income revealed, ex-post utilities assigned Ingredients Village income risk, migration income risk State variables: village income (yj ) and promised utility (w jk ) Choose: migration (I), ex-post utility (h jk ), continuation utility (w j k )

23 Limited commitment constraints Limited commitment constraints: need to receive at least autarky Before migrate: ex-ante After migrate: ex-post Promise keeping constraints Autarky Details of model

24 Intuition

25 Estimation results: with migration Table: Fit of model to data: control Data Model Targeted moments Risk-sharing beta Variance of consumption Mean migration rate Estimated parameters Coeff. relative risk aversion 1.57 Measurement error variance (cons) 0.11 Migration cost 0.06 Set exogenously Discount factor 0.90 Notes: Estimated on data from control villages only.

26 Model with mig: predicts improvement in risk sharing

27 Matching additional experiments Akram et al. (2016) Subsidized different shares of the village Take-up higher when higher share subsidized Risk-sharing: improves; more so in high-share-subsidy Can our model match this Intuition: last figure

28 Conclusion Risk is important in developing countries Context: annual lean season Simple experiment: large increase in migration Large consumption effects But, what spillovers did this have? Examine interaction between risk sharing and migration Townsend: improved risk sharing Structural model: will examine mechanisms further

29

30 Seasonality and Monga Figure 1. Seasonality in Consumption and Price in Rangpur and in Other Regions of Bangladesh Source: Bangladesh Bureau of Statistics 2005 Household Income and Expenditure Survey Return to presentation

31 Summary stats Round 1 Round 4 Roun mean/sd Total Control Treatment Total Control Treatment Total Contr Total income (15.95) (16.42) (15.73) (25.32) (25.44) (26.66) (41.47) (41.14 Wage income (12.06) (13.21) (11.48) (20.62) (20.66) (22.06) (39.25) (50.36 Total consumption (17.12) (17.39) (17.00) (33.80) (33.68) (34.84) (36.91) (34.37 Food consumption (13.37) (13.50) (13.32) (21.74) (21.62) (22.13) (19.34) (19.60 Non-food consumption (5.83) (5.89) (5.81) (16.39) (15.62) (16.99) (23.52) (21.53 Daily per capita calories (0.51) (0.50) (0.51) (0.64) (0.62) (0.65) (0.66) (0.65 Household size (1.30) (1.35) (1.27) (1.43) (1.47) (1.48) (1.46) (1.45 Migrant household (0.49) (0.48) (0.50) (0.49) (0.46 Number of households Return to presentation

32 Consumption effects Table 3: Effects of Migration before December 2008 on Consumption Amongst Remai Panel A: 2008 Consumption Consumption of Food Consumption of Non-Food Total Consumption Total Calories (per person per day) Panel B: 2009 Consumption Consumption of Food Return to presentation Consumption of Non-Food Total Consumption ITT Cash Credit Info ITT ** * ** (29.048) (28.099) (40.177) (24.139) *** ** ** (13.111) (12.425) (17.551) (9.662) *** ** ** (34.610) (33.646) (50.975) (30.593) * ** (62.974) (59.597) (76.337) (47.196) ** (23.076) (23.013) (29.087) (17.589) * * (16.186) (18.135) (20.024) (11.954) ** (30.728) Migration (34.057) and Insurance (39.781) (23.958)

33 Effect of experiment on consumption and income Log Income Log Consumption R5 Control R5 Treatment Back to presentation

34 Migrant sample Table: Treatment effect on financial assistance from and to others, migrant sample Would help you Would help you Would ask you for help Would ask you for help and you d ask and you d help Family *** 0.139*** (0.037) (0.038) (0.044) (0.042) Control mean [0.729] [0.714] [0.497] [0.462] Friends 0.124*** 0.107** 0.127*** 0.106** (0.046) (0.047) (0.041) (0.041) Control mean [0.322] [0.312] [0.266] [0.246] Other villagers 0.096** 0.081** 0.138*** 0.121*** (0.039) (0.041) (0.042) (0.039) Control mean [0.568] [0.518] [0.327] [0.266] NGOs 0.105** 0.112*** (0.041) (0.040) Control mean [0.538] [0.497] Moneylenders (0.030) (0.029) Control mean [0.191] [0.171] Back to presentation

35 Non-migrant sample Table: Treatment effect on financial assistance from and to others, non-migrant sample Would help you Would help you Would ask you for help Would ask you for help and you d ask and you d help Family ** 0.080** (0.029) (0.029) (0.037) (0.035) Control mean [0.715] [0.689] [0.511] [0.466] Friends ** 0.072** (0.032) (0.030) (0.032) (0.029) Control mean [0.260] [0.234] [0.223] [0.195] Other villagers * 0.082** 0.095*** (0.034) (0.034) (0.036) (0.028) Control mean [0.619] [0.573] [0.331] [0.271] NGOs (0.035) (0.033) Control mean [0.582] [0.531] Moneylenders (0.023) (0.022) Control mean [0.181] [0.158] Back to presentation

36 Savings Table: Treatment effect on savings Everyone Migrant sample Non-migrant sample Any Amount Any Amount Any Amount Treatment (0.034) (24.9) (0.049) (37.1) (0.041) (33.9) Control mean N Return to presentation

37 Income moment conditions cov( y i,t, y i,t ) = 2(1 ρ) 1 ρ 2 var(ν) + 2var(ɛy ) (1 ρ)2 cov( y i,t, y i,t 1 ) = 1 ρ 2 var(ν) var(ɛy ) cov( y i,t, y i,t ) = (1 ρ) 1 ρ 2 var(ν) + var(ɛy ) (1 ρ) cov( y i,t, y i,t 1 ) = 1 ρ 2 var(ν) var(ɛy ) ρ(1 ρ) cov( y i,t, y i,t 2 ) = 1 ρ 2 var(ν) ρ(1 ρ) cov( y i,t, y i,t+1 ) = 1 ρ 2 var(ν) Back to presentation

38 Table: Village insurance estimates Control Treatment Difference Income variances Persistent shocks Idiosyncratic (0.015) (0.015) (0.019) Village-aggregate (0.002) (0.004) (0.005) Persistence (0.247) (0.242) (0.335) Transitory shocks Idiosyncratic (0.064) (0.065) (0.036) Village-aggregate (0.007) (0.003) (0.007) Measurement error (0.066) Consumption parameters Persistent shock transmissions Idiosyncratic (0.509) (0.128) (0.511) Village-aggregate (0.694) (0.168) (0.716) Transitory shocks transmissions Idiosyncratic (0.076) (0.044) (0.066) Village-aggregate (0.454) (0.228) (0.534) Measurement error variance (0.008) (0.008) (0.013) Back to presentation

39 Details of autarky Before migration: choose best migration I Ω(y) = max {E ym u((1 I)y + Iy m ) di} + βe y Ω(y ) I After migration and y m realized: Ω(y, I, y m ) = u((1 I)y + Iy m ) di + βe y Ω(y ) Note: Ω(y) = E ym Ω(y, I, y m ) Return to presentation

40 No network effects in migration Table 6. Learning from Own Experience and Others' Experiences in 2009 Re-migration Decision Dep. Var.: Migration in 2009 OLS IV OLS IV OLS IV OLS IV Did any member of the household migrate in 2008? Number of friends and relatives who migrated Number of friends who migrated Number of relatives who migrated Constant 0.392*** 0.410*** 0.392*** 0.464*** 0.393*** 0.436*** 0.392*** 0.476*** (0.02) (0.145) (0.02) (0.133) (0.021) (0.132) (0.02) (0.13) (0.01) (0.022) (0.025) (0.049) (0.011) (0.027) 0.097*** ** *** ** (0.037) (0.083) (0.038) (0.078) (0.037) (0.076) (0.038) (0.077) Observations R-squared *** p<0.01, ** p<0.05, * p<0.1 Robust standard errors in parentheses. Return to presentation

41 Model where V mig = min k { } V (s j ) = min V mig (s), V no mig (s) q π k [ ( 1 1 R V no mig = min [ ( 1 1 R q = (I, h 0, h k, w mig k j 0 s j = (y j, w j0, w j1,..., w jk ), w mig k j k ) C(h jk ) + 1 R ) C(h j0 ) + 1 R j migk τ j V(y j, wj 0, w mig k j,..., w mig k 1 j K ) ] no mig τ j V(y j, w j j 0, w no mig j 1,..., w no mig j K ) ], w no mig j 0, w no mig j k ) k 1,...K, j 1,...J, k 1,..., K

42 Model, cont. (1) ex-post participation constraints: (2) ex-ante participation constraints: Ω(y j k, I) w mig k j k k, j, k Ω(y j, O) w mig k j 0 k, j Ω(y j k, I) w no mig j k j, k Ω(y j, O) w no mig j 0 j ˆΩ(y j ) I k ˆΩ(y j ) I k π k w no mig j k + (1 I )w no mig j 0 j π k w mig k j k + (1 I )w mig k j 0 k, j (3) promise-keeping: w j0 =(1 β)h j0 β τ j (I no mig π k w j k j k w jk =(1 β)h jk β j τ j (I k π k w mig k j k + (1 I no mig )wj 0 ) + (1 I )w mig k j 0 ) k Back to presentation

43 Did migrants have job lead? Table 7. Differences in Characteristics Between Migrants in Treatment and in Control Group Panel A: Percentage of Migrants that Know Someone at Destination Incentive Non incentive Diff First Episode 47% 64% 17*** (1.85) (3.30) (3.8) Any Episode 55% 62% 6.3* (1.80) (3.23) (3.70) Panel B: Percentage of Migrants that had a Job Lead at Destination Incentive Non incentive Diff First Episode 27% 44% 17*** (1.64) (3.41) (3.55) Any Episode 31% 44% 12.8*** (1.67) (3.30) (3.56) Panel C: Percentage of Migrants Traveling Alone Incentive Non incentive Diff First Episode 30% 32% 1.6 (1.70) (3.20) (3.6) Any Episode 37% 37% 0.44 (1.75) (3.20) (3.65) *** p<0.01, ** p<0.05, * p<0.1. Standard errors are in parentheses. Return to presentation

44 People who did well are those who remigrated Figure 4. Migration Experience in 2008 by re-migration Status in Distribution of Total Earnings Density Incentivized Density Not Incentivized In Taka In Taka People who chose to remigrate People who chose not to remigrate People who chose to remigrate People who chose not to remigrate Total Earnings less than Total Earnings less than Back to presentation

45 Estimated income variances Table: Income parameter estimates Control Treatment Difference Persistent shock variance (0.041) (0.072) (0.072) Persistence (ρ) (0.089) (0.125) (0.142) Measurement error variance (0.051) (0.075) (0.080)

46 Estimation of LC model Set β = 0.9 Use income process for control villages Estimate Coefficient of relative risk aversion Variance of measurement error in consumption Moments Risk-sharing beta Observed variance of consumption

47 Estimation results Table: Fit of model to data: control Data Model Targeted moments Risk-sharing beta Variance of consumption Estimated parameters Estimated coeff. relative risk aversion 1.12 Estimated measurement error variance (cons) 0.07 Notes: Estimated on data from control villages only.

48 Model predicts improvement in risk sharing Table: Out-of-sample predictions: treatment Data Model Risk-sharing beta Variance of consumption Estimated coeff. relative risk aversion 1.12 Estimated measurement error variance (cons) 0.07 Notes: Out-of-sample predictions on treatment villages, using parameters estimated on data from control villages only.

49 Mechanics of the model: comparative statics Back to model